Reflection and Reconstruction of Copyright Infringement Logic for AIGC Service Providers

“Ultraman” is a globally recognized anime character. Tsuburaya Productions Co., Ltd. holds the copyright for the Ultraman series and has registered the copyright for various series of Ultraman images. In 2019, Tsuburaya Productions Co., Ltd. signed a “Certificate of Authorization” with the plaintiff, granting exclusive rights to the Ultraman series images to the plaintiff and granting the plaintiff the right to protect those rights.

The website Tab (a pseudonym) is an AI platform operated by the defendant, providing users with AI chat and AI-generated painting services. The plaintiff contends that after users recharge to obtain a Tab membership, they can command Tab to generate images of Ultraman, and the images generated by Tab are substantially similar to the Ultraman images authorized to the plaintiff. Therefore, the defendant’s unauthorized use of the plaintiff’s copyrighted works to train its large model and generate substantially similar images, while profiting from membership recharge and computing power purchase services, constitutes copyright infringement and should bear liability for infringement.

After hearing the case, the Guangzhou Internet Court ruled that the defendant infringed upon the copyright held by the plaintiff and should compensate the plaintiff 10,000 yuan.

Regarding what rights of the plaintiff were infringed by the defendant, the court held that first, the legal fact that Tab generated Ultraman images based on user instructions without permission constituted a reproduction of the original expression of the Ultraman artistic image, infringing the plaintiff’s reproduction rights in the Ultraman works; secondly, some infringing images, such as the fusion of Ultraman with other characters, formed new features while retaining the original expression of Ultraman, infringing the plaintiff’s adaptation rights; finally, for the defendant’s act of generating images and providing them to users, the court stated, “Considering that this case involves new situations of infringement of generative artificial intelligence products, and this court has supported the claims of reproduction rights and adaptation rights infringement, there is no need for repeated evaluation since the alleged infringing act has already been included in the scope of reproduction and adaptation rights control,” thus the court did not find the defendant infringed upon the plaintiff’s rights to disseminate information on the internet.

Regarding whether the defendant should bear compensation liability, the court held: Given that AIGC has certain tool attributes, AIGC service providers should fulfill reasonable duties of care. In this case, the defendant, as a service provider, failed to fulfill reasonable duties of care, thus should bear compensation liability. This is evidenced by the lack of a complaint and reporting mechanism, the absence of potential risk warnings, and the lack of prominent markings. Therefore, the court determined that the defendant should compensate the plaintiff for economic losses of 10,000 yuan.^{[Note 1]}

From a legal perspective, evaluating whether AIGC service providers should bear copyright infringement liability should consider both the input and output ends.

(1) From the Input End

The liability reasons for AIGC service providers primarily focus on the legality and compliance of material acquisition. Currently, there are no specific regulations in China on how to define the act of “legally acquiring materials”. The “Interim Measures for the Management of Generative Artificial Intelligence Services” only provides a general provision stating that “no infringement of others’ legally held intellectual property rights is allowed”^{[Note 2]}. Therefore, according to the current copyright law theoretical system, unless explicit authorization is obtained, AIGC service providers are not allowed to acquire any materials owned by third parties; otherwise, there is a risk of infringing others’ copyright reproduction rights.

(2) From the Output End

The reasons for liability of AIGC service providers mainly concern the potential risk of AIGC-generated content infringing others’ copyrights. Chinese courts generally adopt the standard of “contact + substantial similarity” to evaluate whether copyright infringement has occurred. If this standard is applied, the behavior of AIGC products in capturing materials and conducting deep learning should be regarded as having met the “contact” requirement; as for the “substantial similarity” requirement, current developments in AIGC products have matured, and the generated content increasingly meets users’ personalized needs. As evidenced by the plaintiff in this case, the content generated by AIGC based on user instructions (i.e., “Ultraman”) has legally replicated the original expression of the Ultraman artistic image; thus, the content generated by AIGC based on precise instructions is likely to be considered “substantially similar” to the instructed object, leading AIGC service providers into copyright infringement lawsuits.

It should be noted that according to the copyright infringement liability logic commonly adopted in China’s judicial system, the output content generated by the defendant’s Tab, as a commercial AIGC, constitutes substantial similarity with the artistic works owned by the plaintiff, and based on the high recognition of “Ultraman” in society, even if the plaintiff does not provide direct evidence that the defendant indeed captured the plaintiff’s copyrighted “Ultraman” related images during the earlier material capturing process, it should still be inferred that the defendant has met the “contact” requirement. Therefore, the court’s ruling that the defendant bears infringement liability and compensates relevant damages is legally valid.

According to existing copyright legislation and judicial protection systems, AIGC service providers must ensure that the materials captured for training AIGC products do not infringe others’ copyrights. Therefore, in commercial practice, companies mainly obtain relevant materials through the following methods:

(1) Open Source Datasets

From a legal perspective, open source datasets are databases that are open to the public and licensed for free use, modification, and sharing by copyright holders. Currently, for most small and medium-sized enterprises lacking sufficient economic and technical strength, open source datasets are one of the few reliable sources of data that can mitigate copyright infringement risks to some extent. Overall, AIGC service providers have a relatively low risk of copyright infringement when training AIGC products using open source datasets; however, on the other hand, open source datasets have certain limitations in quantity and quality to maximize copyright infringement risk avoidance. If AIGC service providers solely rely on open source datasets to train AIGC products, the output content may ultimately fail to meet user expectations in terms of quality and may even lead to “algorithmic bias”^{[Note 3]}, which in turn makes it difficult for AIGC service providers to achieve sufficient economic benefits, hindering the development of the AI industry.

(2) Data Purchase

This refers to AIGC service providers acquiring relevant data for training AIGC products by paying compensation (including monetary payment or exchanging their own data, also known as “data exchange”). For AIGC service providers intending to offer differentiated services, purchasing relevant data is essential for differentiating training of AIGC products. However, acquiring relevant data from others through purchase also presents certain disadvantages. First, if AIGC service providers obtain relevant data through purchase, they will inevitably incur high economic costs. For example, OpenAI, the provider of ChatGPT, reportedly pays millions of dollars annually just for copyright authorization of certain media news content^{[Note 4]} to maintain its leading position; in China, for instance, companies generally pay authorization fees ranging from dozens to hundreds of yuan for each image when purchasing images for commercial reasons. For small and medium-sized enterprises with weaker economic strength, the high authorization fees create significant obstacles for AIGC service providers to purchase data for training AIGC products. On the other hand, even if AIGC service providers acquire relevant data for training AIGC products through purchase, if the source of that data is illegal, AIGC service providers will still bear copyright infringement liability. From a practical perspective, even if AIGC service providers have legally purchased relevant data, they still cannot escape the obligation to conduct prior reviews of the data, further increasing the economic burden and legal risks for AIGC service providers.

(3) Obtaining from Public Channels

This refers to AIGC service providers obtaining data for training AIGC products through manual or technical means from the internet or other public channels. AIGC service providers can acquire a large amount of usable data through this method, but they also face the most severe legal risks. On an international level, OpenAI has recently been involved in multiple lawsuits for allegedly illegally acquiring relevant data for training AIGC products^{[Note 5]}, while Stability AI, the provider of stable diffusion, has also been sued by rights holders for scraping images without permission^{[Note 6]}. According to China’s copyright protection system, if relevant data is copyright protected, the behavior of AIGC service providers acquiring data through web scraping is almost certain to infringe the reproduction rights of copyright holders; and due to the commercial nature of AIGC products, it is also difficult for AIGC service providers to raise a fair use defense. If AIGC service providers obtain data through manual or technical means from public channels, they may face administrative, civil, or even criminal liability.

(4) Obtaining from AIGC Product Users

When AIGC product users issue commands for AIGC to generate specific content, they generally need to input text commands or upload relevant images to achieve their goals. The data provided by AIGC product users is one of the important sources from which AIGC service providers obtain data. Although AIGC service providers can formally prohibit AIGC product users from providing text, images, or other content that infringes others’ copyrights through user agreements and risk warnings, similar to the situation of acquiring data through purchase, if the content uploaded by AIGC product users infringes others’ copyrights, AIGC service providers will also bear infringement liability for obtaining the uploaded content.

Overall, according to the current copyright infringement liability logic in China, the conditions for AIGC service providers to legally acquire data for training AIGC products are quite stringent, and they require significant economic and technical costs. While this liability logic can match the current copyright theoretical system, it objectively creates a considerable legal “barrier” for AIGC service providers.

The “Safe Harbor Principle” is a widely adopted principle in the intellectual property protection systems of various countries, and it is also reflected in Article 22 of China’s “Regulations on the Protection of Rights of Information Network Dissemination”^{[Note 7]}. However, under China’s current copyright protection system, it is difficult for AIGC service providers to be included in the “Safe Harbor” for protection. Taking this case as an example, the defendant does not merely provide “storage space” for other service users; rather, it generates images that allegedly infringe copyright based on AIGC users’ instructions and obtains economic benefits by selling membership qualifications. Given that Ultraman is a well-known anime IP, the defendant should have been aware of the potential for copyright infringement in its AIGC products. Although the defendant promptly deleted the related content as requested by the plaintiff, it still needs to bear related liability. Without the protection of “Safe Harbor”, if we follow the commonly adopted standard of “contact + substantial similarity” for determining copyright infringement in China, to minimize potential copyright infringement risks, AIGC service providers typically need to take measures in the following two aspects:

(1) Avoid Using Materials That May Infringe Others’ Copyrights at the Input End

This approach aims to sever the “contact” element in the copyright infringement determination standard. For AIGC products, if no materials that may infringe others’ copyrights are used at the input end, it may not legally constitute “contact”, thereby mitigating copyright infringement risks to some extent. However, according to the working principles of AIGC, the content at the input and output ends is not one-to-one corresponding. Therefore, even if no materials that may infringe others’ copyrights are used at the input end, at the output end, it is still possible to generate works that are substantially similar to copyrighted works based on user instructions and prompts, potentially leading to copyright infringement lawsuits; at the same time, for certain well-known IPs (like “Ultraman” in this case), even if the AIGC input end did not use related materials, the court may still determine that the AIGC service provider should have had contact with the copyright holder’s works, thus holding the AIGC service provider liable.

(2) Conduct Compliance Control at the Output End

This means that AIGC service providers should design AIGC products to include compliance algorithms at the output end, allowing the AIGC product to conduct compliance self-checks on the content to be generated after receiving user instructions, ensuring that the generated content does not infringe others’ copyrights. Currently, this method is widely adopted by AIGC service providers and is an effective measure to mitigate risks. However, based on the actual working conditions of AIGC, it is often difficult to completely filter out all potentially infringing content using compliance algorithms. As seen in this case, even though the defendant deleted the allegedly infringing content and made technical adjustments, the court could still generate images similar to Ultraman by inputting other keywords related to Ultraman. On the other hand, given the efficiency of AIGC and market demand, it is also impractical to conduct manual reviews of all generated content. Furthermore, due to the ambiguity in evaluating whether works are substantially similar, requiring AIGC service providers to completely eliminate the possibility of copyright infringement in all generated content is technically unfeasible and legally unreasonable, thus unduly increasing the liability of AIGC service providers.

As mentioned earlier, due to the unique technical characteristics and legal nature of AIGC products, AIGC service providers lack sufficient possibilities for exemption under China’s current copyright protection system. This leads to AIGC service providers being easily caught in legal risks concerning copyright, which also negatively affects the development of China’s AI industry.

In fact, at the legislative level, from the “Interim Measures for the Management of Generative Artificial Intelligence Services” issued on July 10, 2023, by the National Internet Information Office, the National Development and Reform Commission, the Ministry of Education, the Ministry of Science and Technology, the Ministry of Industry and Information Technology, the Ministry of Public Security, and the National Radio and Television Administration, to the “Basic Requirements for the Safety of Generative Artificial Intelligence Services” issued by the National Information Security Standardization Technical Committee on March 1, 2024, we can observe that the state is gradually loosening legal restrictions on AIGC service providers. As stated in Article 8 of the “Basic Requirements for the Safety of Generative Artificial Intelligence Services”, due to the technical characteristics of AIGC products, specific requirements for the overall qualification rate of AIGC-generated content should be established, rather than demanding that all content generated by AIGC be free of illegal or irregular situations. At the judicial level, the Guangzhou Internet Court, in discussing the reasoning behind the defendant’s liability for infringement, did not solely adopt the “contact + substantial similarity” method but emphasized that the defendant’s liability for compensation was due to its failure to fulfill reasonable duties of care. In fact, the “duty of care” required of the defendant by the Guangzhou Internet Court includes establishing a complaint and reporting mechanism, increasing potential risk warnings, and providing prominent markings, all of which are achievable for AIGC service providers. If China’s judicial system adopts the ruling perspective of the Guangzhou Internet Court in this case, it can be believed that a brand new “safe harbor rule” applicable to AIGC service providers can be successfully established, providing sufficient legal protection for AIGC service providers.

As AIGC technology continues to develop and be widely applied, to adapt to the new social landscape, I believe that the copyright infringement liability logic for current AIGC service providers should be reconstructed from both ends based on the technical characteristics of AIGC products, ultimately achieving the goal of promoting industrial upgrading and social development while balancing the interests of all parties.

1. At the input end, akin to Japan’s current AI regulatory model, allow AIGC to broadly capture data that can be obtained from public channels and does not involve personal information or commercial secrets, and that the behavior of AIGC service providers in capturing data does not constitute copyright infringement unless it seriously impacts the actual interests of the rights holders.

2. At the output end, similar to the “safe harbor principle” in traditional copyright protection systems, allow AIGC service providers without subjective malice to exempt themselves from copyright infringement liability through a notice-and-takedown model.

3. Meanwhile, establish a copyright licensing department for AIGC coordinated by the state, coordinating copyright holders to provide unified licenses to AIGC service providers. In terms of fees, the mechanism can refer to the authorization licensing mechanism of music software, where copyright holders provide copyright licenses to AIGC service providers, who then pay the relevant copyright authorization fees to copyright holders and subsequently charge related fees to AIGC users through membership recharge or separate charges, ultimately allowing all parties to achieve reasonable benefits.

AIGC, as an important technological form in the current development of the AI industry, poses unprecedented challenges to the existing copyright protection system. In response, adjustments should be made to the copyright protection system and liability logic in a timely manner to adapt to social development and industrial upgrading, providing a safeguard for the transformative wave brought by AIGC technology.